SimPseudo: Enhancing whole slide imaging classification with similarity and prototype-based instance-level pseudo-supervision

Whole Slide Imaging (WSI) analysis in digital pathology is challenged by ultra-high resolution and limited pixel-level annotations. Multiple Instance Learning (MIL) mitigates these issues by relying on image-level labels; however, conventional Top-K pseudo labeling strategies supervise only high-confidence instances, leaving the majority unlabeled. To address this limitation, we propose SimPseudo, a similarity-based pseudo label generation framework that assigns reliable soft labels to all WSI instances. SimPseudo comprises three major components: a Top-N₁ similarity strategy that integrates MIL classifier confidence with pre-trained negative prototypes to generate adaptive soft pseudo labels, a negative prototype generation mechanism that constructs representative prototypes from true negatives to improve instance labeling, and a pseudo label correction module that refines and stabilizes labels using classifier-derived probabilities. The framework is model-agnostic and can be seamlessly integrated into diverse MIL classifiers. Extensive experiments on Camelyon16 and Camelyon17 demonstrate consistent gains over Top-K baselines. In particular, SimPseudo achieves improvements of up to 1.9 % in accuracy and 4.54 % in AUC on Camelyon16, and further yields 1.4 % accuracy and 0.8 % AUC improvements on Camelyon17. These results demonstrate the effectiveness of SimPseudo in strengthening instance-level supervision and enhancing WSI-based pathology analysis.